Although Nvidia is facing multiple pressures near the end of the year, according to supply chain sources, Nvidia still plans to release its next-generation GB300 AI server platform at the GTC conference to be held from March 17 to 20, 2025.
The platform has undergone a series of new upgrades, achieving an all-round improvement in performance and configuration, and will become a new tool for NVIDIA to increase its performance. Moreover, multiple improvements are aimed at elevating performance and efficiency to further meet the growing needs of AI computing.
NVIDIA GB300 server fully upgraded
Specifically, NVIDIA GB300 AI server integrates the new Blackwell Ultra series B300 GPU, with a power of 1400W and a single-card FP4 performance improvement of 1.5 times.
At the HBM level, compared with the eight-layer stack of GB200, GB300 will use HBM3E with a 12-layer stack, and the capacity will be increased from 192GB to 288GB. It will also introduce new liquid cooling trays, cold plates and UQD (Universal Quick Disconnect) parts for liquid cooling lines.
In addition, the optical network card will be upgraded from CX7 to CX8, and optical transceiver will be upgraded from 800G to 1.6T.
Other upgrades include the use of a slot design, the computing board will use LPCAMM, and the capacitance tray may become standard for the GB300 NVL72 cabinet, and the BBU (battery backup unit) may be optional.
The supply chain pointed out that the mass production price of a BBU module is about $300, and it is estimated that the total price of the BBU in a GB300 system will reach $1,500; The production price of a supercapacitor is about US$20 to US$25, and a GB300 NVL72 rack requires more than 300 capacitance.
However, there have been recent reports that Nvidia encountered supply chain challenges in the development of the GB300 and B300 AI servers, especially the overheating problem of the key component 5×5 DrMOS chip produced by AOS, which may affect the mass production progress.
Foxconn will be the biggest winner
In response to Nvidia’s product upgrades, supply chain manufacturers such as Foxconn and Quanta have acted quickly and entered the R&D and design stage of the GB300 AI server.
It is reported that NVIDIA has preliminarily finalized the configuration of the NVIDIA GB300 AI server order, with Foxconn still being the largest supplier. NVIDIA GB300 AI servers are expected to be available in the first half of 2025, leading its global peers.
Industry insiders pointed out that just like Apple’s strategy of launching a new iPhone every year, Nvidia CEO Jensen Huang has previously announced that Nvidia will also follow the pace of “launching a new generation of AI products every year.” Major partners must deploy in advance to keep up with Nvidia’s product development pace.
Supply chain analysis shows that in addition to its highly vertically integrated advantages, Foxconn Group has also invested heavily in NVIDIA’s early research and development of AI servers. The two parties began cooperating as early as 2017, and have established a close cooperative relationship on the development of new products from the first generation of AI servers to the current GB200 and even GB300.
Taking the GB200 AI server as an example, in addition to the GPU and CPU, the components that Foxconn can provide account for about 80% to 90%.
Industry insiders revealed that Quanta and Inventec are also important partners of NVIDIA’s GB300 AI server. In terms of order share, Quanta is the second largest supplier after Foxconn, while Inventec is expected to grow.
GB200 will see peak shipment in 2025
According to the latest market research, Nvidia’s GB200 chip shipments in the fourth quarter of 2024 are expected to be approximately 150,000 to 200,000 units, and shipments in the first quarter of 2025 are expected to increase significantly by 200% to 250%.
However, the shipment schedule of GB200 cabinets has been delayed. Since the solution’s design specifications in terms of high-speed interconnect interface and thermal design power consumption are significantly higher than those of the mainstream market, the supply chain needs more time for adjustment and optimization. It is expected that the peak shipment of GB200 cabinets will be delayed to between the second and third quarters of 2025.
In addition, the TDP value of the GB200 NVL72 cabinet is as high as 140KW, and the traditional air cooling solution can no longer meet its needs. Therefore, liquid cooling technology has become a necessity, and related industries are stepping up their research and development efforts on liquid cooling components to improve heat dissipation efficiency.
Overall, although the shipment volume of GB200 chips is growing rapidly, it will still take time for the cabinet-type solution to be fully rolled out, and it is expected that the shipment peak will come in mid-2025.
Looking further, GB300’s high power consumption design is expected to promote the development of the liquid cooling industry, from which related supply chain manufacturers may benefit.